Evaluation of compact deep neural network models for chest x-ray image classification

Abstract

Detection of abnormalities in chest X-ray images is primarily challenging due to time constraints and the excessive workload of medical professionals. The uneven distribution and occasionally unreliable labels further complicate the reliable construction of automated systems based on neural networks. In this paper, three compact neural network architectures are investigated: MobileNetV4, EfficientFormerV2, and EdgeNeXt, and their performance is compared. Using binary classification and focal loss, adaptation to each abnormality is implemented separately, thus enabling a different approach than the more commonly used multi-label method. The models were trained on the NIH ChestX-ray14 dataset with three-fold stratified cross-validation and minimal augmentation in favor of preserving image structure. EdgeNext and MobileNetV4 achieve competitive performance compared to larger, traditional networks, with average ROC AUC values of 79.59% and 77.73%, while EfficientFormerV2 achieves 70.95%. This confirms the potential of efficient models for application in clinical settings and opens up the possibility of integrating automatic report generation and model explanation.